Ignore repeated prompt

MANIFEST.in

@@ -1,4 +1,3 @@
 include faster_whisper/assets/silero_vad.onnx
 include requirements.txt
 include requirements.conversion.txt
-include faster_whisper/assets/pyannote_vad_model.bin

README.md

@@ -69,6 +69,7 @@ segments, info = model.transcribe("audio.mp3", beam_size=5, language="en")
 
 * Python 3.8 or greater
 
+Unlike openai-whisper, FFmpeg does **not** need to be installed on the system. The audio is decoded with the Python library [PyAV](https://github.com/PyAV-Org/PyAV) which bundles the FFmpeg libraries in its package.
 
 ### GPU
 
@@ -165,35 +166,6 @@ for segment in segments:
 segments, _ = model.transcribe("audio.mp3")
 segments = list(segments)  # The transcription will actually run here.
 ```
-
-### multi-segment language detection
-
-To directly use the model for improved language detection, the following code snippet can be used:
-
-```python
-from faster_whisper import WhisperModel
-model = WhisperModel("medium", device="cuda", compute_type="float16")
-language_info = model.detect_language_multi_segment("audio.mp3")
-```
-
-### Batched faster-whisper
-
-
-The batched version of faster-whisper is inspired by [whisper-x](https://github.com/m-bain/whisperX) licensed under the BSD-2 Clause license and integrates its VAD model to this library. We modify this implementation and also replaced the feature extraction with a faster torch-based implementation. Batched version improves the speed upto 10-12x compared to openAI implementation and 3-4x compared to the sequential faster_whisper version. It works by transcribing semantically meaningful audio chunks as batches leading to faster inference. 
-
-The following code snippet illustrates how to run inference with batched version on an example audio file. Please also refer to the test scripts of batched faster whisper.
-
-```python
-from faster_whisper import WhisperModel, BatchedInferencePipeline
-
-model = WhisperModel("medium", device="cuda", compute_type="float16")
-batched_model = BatchedInferencePipeline(model=model)
-segments, info = batched_model.transcribe("audio.mp3", batch_size=16)
-
-for segment in segments:
-    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
-```
-
 ### Faster Distil-Whisper
 
 The Distil-Whisper checkpoints are compatible with the Faster-Whisper package. In particular, the latest [distil-large-v3](https://huggingface.co/distil-whisper/distil-large-v3)

benchmark/wer_benchmark.py

@@ -1,6 +1,5 @@
 import argparse
 import json
-import os
 
 from datasets import load_dataset
 from evaluate import load
@@ -27,9 +26,7 @@
 
 # define the evaluation metric
 wer_metric = load("wer")
-
-with open(os.path.join(os.path.dirname(__file__), "normalizer.json"), "r") as f:
-    normalizer = EnglishTextNormalizer(json.load(f))
+normalizer = EnglishTextNormalizer(json.load(open("normalizer.json")))
 
 
 def inference(batch):

docker/Dockerfile

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04
+FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
 WORKDIR /root
 RUN apt-get update -y && apt-get install -y python3-pip
 COPY infer.py jfk.flac ./

faster_whisper/__init__.py

@@ -1,13 +1,12 @@
 from faster_whisper.audio import decode_audio
-from faster_whisper.transcribe import BatchedInferencePipeline, WhisperModel
+from faster_whisper.transcribe import WhisperModel
 from faster_whisper.utils import available_models, download_model, format_timestamp
 from faster_whisper.version import __version__
 
 __all__ = [
     "available_models",
     "decode_audio",
     "WhisperModel",
-    "BatchedInferencePipeline",
     "download_model",
     "format_timestamp",
     "__version__",

faster_whisper/audio.py

@@ -1,7 +1,19 @@
+"""We use the PyAV library to decode the audio: https://github.com/PyAV-Org/PyAV
+
+The advantage of PyAV is that it bundles the FFmpeg libraries so there is no additional
+system dependencies. FFmpeg does not need to be installed on the system.
+
+However, the API is quite low-level so we need to manipulate audio frames directly.
+"""
+
+import gc
+import io
+import itertools
+
 from typing import BinaryIO, Union
 
-import torch
-import torchaudio
+import av
+import numpy as np
 
 
 def decode_audio(
@@ -17,42 +29,91 @@ def decode_audio(
       split_stereo: Return separate left and right channels.
 
     Returns:
-      A float32 Torch Tensor.
+      A float32 Numpy array.
 
       If `split_stereo` is enabled, the function returns a 2-tuple with the
       separated left and right channels.
     """
+    resampler = av.audio.resampler.AudioResampler(
+        format="s16",
+        layout="mono" if not split_stereo else "stereo",
+        rate=sampling_rate,
+    )
+
+    raw_buffer = io.BytesIO()
+    dtype = None
 
-    waveform, audio_sf = torchaudio.load(input_file)  # waveform: channels X T
+    with av.open(input_file, mode="r", metadata_errors="ignore") as container:
+        frames = container.decode(audio=0)
+        frames = _ignore_invalid_frames(frames)
+        frames = _group_frames(frames, 500000)
+        frames = _resample_frames(frames, resampler)
+
+        for frame in frames:
+            array = frame.to_ndarray()
+            dtype = array.dtype
+            raw_buffer.write(array)
+
+    # It appears that some objects related to the resampler are not freed
+    # unless the garbage collector is manually run.
+    del resampler
+    gc.collect()
+
+    audio = np.frombuffer(raw_buffer.getbuffer(), dtype=dtype)
+
+    # Convert s16 back to f32.
+    audio = audio.astype(np.float32) / 32768.0
 
-    if audio_sf != sampling_rate:
-        waveform = torchaudio.functional.resample(
-            waveform, orig_freq=audio_sf, new_freq=sampling_rate
-        )
     if split_stereo:
-        return waveform[0], waveform[1]
+        left_channel = audio[0::2]
+        right_channel = audio[1::2]
+        return left_channel, right_channel
+
+    return audio
+
+
+def _ignore_invalid_frames(frames):
+    iterator = iter(frames)
+
+    while True:
+        try:
+            yield next(iterator)
+        except StopIteration:
+            break
+        except av.error.InvalidDataError:
+            continue
+
+
+def _group_frames(frames, num_samples=None):
+    fifo = av.audio.fifo.AudioFifo()
+
+    for frame in frames:
+        frame.pts = None  # Ignore timestamp check.
+        fifo.write(frame)
+
+        if num_samples is not None and fifo.samples >= num_samples:
+            yield fifo.read()
+
+    if fifo.samples > 0:
+        yield fifo.read()
+
 
-    return waveform.mean(0)
+def _resample_frames(frames, resampler):
+    # Add None to flush the resampler.
+    for frame in itertools.chain(frames, [None]):
+        yield from resampler.resample(frame)
 
 
 def pad_or_trim(array, length: int, *, axis: int = -1):
     """
     Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
     """
-    axis = axis % array.ndim
     if array.shape[axis] > length:
-        idx = [Ellipsis] * axis + [slice(length)] + [Ellipsis] * (array.ndim - axis - 1)
-        return array[idx]
+        array = array.take(indices=range(length), axis=axis)
 
     if array.shape[axis] < length:
-        pad_widths = (
-            [
-                0,
-            ]
-            * array.ndim
-            * 2
-        )
-        pad_widths[2 * axis] = length - array.shape[axis]
-        array = torch.nn.functional.pad(array, tuple(pad_widths[::-1]))
+        pad_widths = [(0, 0)] * array.ndim
+        pad_widths[axis] = (0, length - array.shape[axis])
+        array = np.pad(array, pad_widths)
 
     return array